腾讯的 AI 故事,正在进入一个有些沉重的阶段。
11 月 18 日,腾讯发布 2026 年第三季度财报。本季度营收 2073.4 亿元,同比增长 8%;Non-IFRS 经营利润 681 亿元,同比下降 6%;Non-IFRS 归母净利润 665 亿元,同比下降 6%。
本季度,腾讯资本开支达到 814 亿元,同比增长 527%,单季支出已经超过 2025 年全年;研发投入 323.6 亿元,同比增长 42%。大规模算力采购和预付款之下,自由现金流为负 691 亿元;若剔除算力采购预付款,自由现金流仍有 169 亿元。
单看收入,这依旧是一份稳健的财报。但把 AI 拆出来,曲线的分叉比上一季度更明显:若剔除混元、元宝、CodeBuddy、WorkBuddy(了解更多)、小微等新 AI 产品的收入、成本与开支,腾讯本季度 Non-IFRS 经营利润为 833 亿元,同比增长 8%,经营利润率 40%,和去年同期持平。
也就是说,老业务的盈利能力没有出问题,但新 AI 业务在这个季度对 Non-IFRS 经营利润造成了约 152 亿元的净影响,上一季度这个数字是 105 亿元。整体利润增速,就这样被拉到了负值区间。
腾讯把越来越多的当期现金,换成了 GPU、模型、Agent 和下一代入口。
一、「AI 折价」:152 亿口径差,和 814 亿的决心
三季度,腾讯营收 2073.36 亿元,同比增长 8%,环比增长 1%。三大板块里,营销服务收入 430 亿元,同比增长 19%,是增速最快的板块;金融科技及企业服务收入 637 亿元,增长 9%;增值服务收入 983 亿元,增长 3%。微信及 WeChat 合并月活账户数达到 14.49 亿。
利润端的数字就没那么好看了。按非国际财务报告准则,腾讯三季度经营利润 681 亿元,同比下降 6%;IFRS 口径下经营利润 591 亿元,同比下降 7%。Non-IFRS 净利润 665 亿元,同样下滑 6%。
若剔除新 AI 产品的贡献,三季度 Non-IFRS 经营利润为 833 亿元,同比增长 8%,经营利润率 40%,与去年同期持平。两个口径之间相差约 152 亿元。「新 AI 产品」,主要包括混元、元宝、CodeBuddy、WorkBuddy 和小微。
也就是说,拖累利润的是正在投入期的 AI 业务。
三季度腾讯资本开支总额达到 814 亿元,同比增长 527%,单季支出已经超过 2025 年全年;前三季度累计资本开支 1662 亿元,同比增长 179%,钱主要流向 IT 基础设施、数据中心和土地使用权。现金口径的资本开支付款为 651 亿元,同比增长 226%。研发投入 323.6 亿元,同比增长 42%;一般及行政开支同比增长 29% 至 441 亿元,若剔除研发开支,这项费用仅增长 2%,非 AI 开支的管控依然很紧。
三季度自由现金流为负 691 亿元。经营活动产生的现金净额只有 26 亿元,而资本开支、媒体内容、租赁负债三笔付款合计 717 亿元。腾讯解释,经营现金流里包含大额算力采购预付款,主要用于混元升级、WorkBuddy 推理和云客户需求;若剔除这笔预付款,自由现金流本应为正 169 亿元。
资金状况随之改变,6 月底还有 582 亿元净现金,9 月底变为 137 亿元净债务。官方解释,这主要由大额算力采购造成,投资组合中处置超过新增投资带来的 84 亿元净流入,只抵消了一部分。腾讯账上仍有 5257 亿元现金及定期存款,持有的上市、非上市投资合计约 9026 亿元。
股东回报方面,腾讯三季度回购约 1587 万股、耗资约 70 亿港元,回购股份已注销。
市场对此有不小分歧,多家券商在财报前下调了目标价,野村、瑞银等机构早早提示 AI 支出对利润率的压制;但花旗、美银、里昂等多数大行仍维持「买入」评级,核心理由是短期利润承压夯实的是长期竞争力。这本质上是一道时间题:投资者愿意等多久,取决于回款来得有多快。
二、年化 5 亿美元的 WorkBuddy 与削减支出的元宝
电话会上,刘炽平披露了 WorkBuddy 的收入数据,按过去七天的收入折算,WorkBuddy 年化收入已经超过 5 亿美元,是腾讯会议营收的两倍以上;三季度单季收入环比二季度增长超过 5 倍。
用户侧,WorkBuddy 日活突破 1000 万,每天消耗超过 100 万亿 token,按大语言模型 token 消耗量计,已是国内第一的 AI 应用。
企业端的渗透上,企业版今年 6 月推出,中国百强企业里已有 60% 成为客户,中国企业 500 强覆盖率达到 50%,落地 50 多个行业。
关于这套变现路径,刘炽平把 WorkBuddy 类比为网络游戏的演进:订阅加 token 用量的复合付费,让用户以低门槛进入,随着获得价值增加而扩大支出,「当消费可以沿需求曲线向上扩展,就会解锁一个大得多的收入池」。针对市场担心轻量模型 flash 会拉低 token 定价,他的回应是,flash 是获客工具,更强的模型负责价值变现,两者刚好覆盖需求曲线上的不同位置。
模型底座上,今年以来混元保持约两个月一次大版本迭代。11 月 4 日发布的 Hy4,发布时位列 Agent Arena 中国模型第一、全球开源模型第一,幻觉率从 12.95% 降到 7.2%,首次引入视觉理解和计算机操作能力,重点强化编程、企业工作流和科研这些高价值场景。
本周一推出的 Hy4 flash,总参数 224B、激活参数仅 12B,单任务成本约为 Hy4 的二十分之一,上线首日 token 调用量突破 50 万亿,登顶 OpenRouter 日榜。Hy4 在 WorkBuddy 首发当天,用户额度大量触顶,随即完成 10 倍扩容,模型和产品之间的协同已经相当紧。
云业务方面,三季度云服务收入同比增速从上季度的 21% 跳升到 35%,其中模型即服务收入环比增长超过三倍,GPU 租赁收入同比增长超 30%,WorkBuddy 和 MaaS 收入都实现了环比三倍以上。管理层预计,这个较高增速会贯穿 2026 年全年。
C 端的想象空间则在微信里。刘炽平披露,微信内嵌的 AI 助手小微,日活已经达到 3000 万,下一步会接入朋友圈、公众号、视频号等 30 余个触点。
他给出了一个成本数据,即便小微向全部 14 亿微信用户开放,单季度运营成本仍低于元宝今年年初的单季度成本,因为它跑在专为低成本定制的轻量模型上。
同时,考虑到算力效率每年成倍提升,预计一年到 18 个月内小微能力大幅跳升,最终达到 Meta 产品的水平,同时覆盖全部微信用户。
商业化方面,小微将成为微信现有收入的放大器,加速收入增长,「在微信现有收入基础上增加 10% 到 20%,就已经是非常可观的数字,未来 6–12 个月路径会更清晰」。广告、交易变现是近期机会,后续还可以新增智能代理撮合服务等收入。
至于元宝,刘炽平则谈到,「元宝没有我们预期的那样成功。」元宝投入不及预期,已大幅削减支出,转向能力建设并上线 AI 搜索。
当然,5 亿美元是七天收入的年化折算,绝对值和腾讯整体体量相比仍小,AI 业务整体还处在投入期。这些新增收入能否覆盖成本,还需要后面几个季度持续验证。但至少,「模型— 产品— 收入」的飞轮,已经在转动。
财报数字之外,这场转型最难的部分,腾讯管理层在电话会和内部讲话里都没有回避。
首先是算力的硬约束。面对分析师追问,刘炽平承认,无论是租赁还是自建算力都面临约束,模型训练不可避免会受到影响,短期内下一代混元的推出可能需要更长时间。
他同时区分了两个环节:推理目前不受影响,且计算效率逐年快速提升;长期看,国内供应链会逐步进入推理环节,国产化进程还会加快。这意味着腾讯要在一个有天花板的算力环境里安排训练节奏,模型迭代的时间表开始受到外部条件的真实牵制。
11 月 11 日腾讯 28 周年司庆当晚,马化腾在内部 LinkTime 上的讲话迅速流出,措辞严厉程度近年罕见。他把当前形势比作「浪已经杀过来」,远处的浪潮正在吞没一个个山丘,「你不能还坐着欣赏」,要求核心业务从「陆战转海战」,一半人像陆军一样守好现有业务,另一半人像海军一样去探索 AI Native 新战场。
如果放在更长的时间线里来看——2023 年马化腾谈 AI 的关键词还是「不着急」,混元发布晚于同行,元宝长期未获推广。今年 1 月他已经在年会上复盘「动作慢了」。随后腾讯启动了一系列密集调整: OpenAI 前研究员姚顺雨出任首席 AI 科学家,统筹混元研发;新设 AI Infra、AI Data 等部门,2026 年 3 月撤销成立十年的 AI Lab 并入混元体系,算力改为总部统一调配;混元团队把汇报线从三级压缩到两级,WXG 试行项目负责制,游戏发行线淡化总监、组长头衔。
电话会上,马化腾谈到两个他最满意的样本:模型团队完全从零重建,一年内把混元推到中国顶尖水平;WorkBuddy 团队六七个月前几乎一无所有,做出了今天的业务。他想要的理想状态是,「如果我们能让现有业务用更少的人运转,把这些人重新部署去构建新的增长驱动力,这就是理想状态。」
这也是一周前腾讯内部讲话的核心逻辑。与上一轮被动降本不同,这次被定义为「主动增效」:组织扁平化,以小团队为作战单元,减少中间传话和审批环节,明确提出「未来不一定要那么多管理岗位」,要打破山头主义和地盘意识。
当 AI 把开发门槛大幅压低,一个带着数字员工的小团队就能完成过去需要整条协作链的工作,传统金字塔式的管理层级,正在失去存在的理由。
财报发布当天,腾讯股价收在 372.8 港元,年内跌幅约 37%。财报落地并未止住跌势,次日早盘股价应声下探,一度跌破 350 港元。
从这一季的数字看,腾讯的 AI 叙事已经告别了过去互联网公司熟悉的轻研发、快迭代,走向重算力、长周期、生态化,814 亿元资本开支和 691 亿元负现金流是当下阶段的一种阵痛。WorkBuddy 能不能长成真正的生产力平台,小微能不能把微信带进 AI 时代,混元能不能在受限的算力里站稳第一梯队,都还要后续几个季度给出答案。
本文来源:https://www.geekpark.net/news/374302
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Tencent's AI story is entering a somewhat heavy phase.
On November 18, Tencent released its 2026 third-quarter financial results. Revenue for the quarter was RMB 207.34 billion, up 8% year over year; Non-IFRS operating profit was RMB 68.1 billion, down 6% year over year; and Non-IFRS net profit attributable to shareholders was RMB 66.5 billion, down 6% year over year.
This quarter, Tencent's capital expenditure reached RMB 81.4 billion, up 527% year over year, with single-quarter spending already exceeding the entire 2025 total; R&D investment was RMB 32.36 billion, up 42% year over year. Amid large-scale computing power purchases and prepayments, free cash flow was negative RMB 69.1 billion; excluding computing power purchase prepayments, free cash flow was still RMB 16.9 billion.
Taken on revenue alone, this is still a solid set of results. But when AI is broken out, the divergence in the curves is more pronounced than last quarter: excluding the revenue, costs, and expenses of new AI products such as Hunyuan, Yuanbao, CodeBuddy, WorkBuddy(了解更多), and Xiaowei, Tencent's Non-IFRS operating profit this quarter was RMB 83.3 billion, up 8% year over year, with an operating margin of 40%, flat with the same period last year.
In other words, there is nothing wrong with the profitability of the legacy businesses, but the new AI business had a net impact of about RMB 15.2 billion on Non-IFRS operating profit this quarter, versus RMB 10.5 billion last quarter. That is how overall profit growth was pulled into negative territory.
Tencent is converting more and more of its current cash into GPUs, models, agents, and next-generation entry points.
I. The 'AI discount': a RMB 15.2 billion accounting gap and RMB 81.4 billion of determination
In the third quarter, Tencent's revenue was RMB 207.336 billion, up 8% year over year and 1% quarter over quarter. Among its three major segments, marketing services revenue was RMB 43 billion, up 19% year over year, the fastest-growing segment; fintech and business services revenue was RMB 63.7 billion, up 9%; and value-added services revenue was RMB 98.3 billion, up 3%. Combined monthly active accounts for Weixin and WeChat reached 1.449 billion.
The profit side looks less attractive. Under non-IFRS, Tencent's third-quarter operating profit was RMB 68.1 billion, down 6% year over year; under IFRS, operating profit was RMB 59.1 billion, down 7%. Non-IFRS net profit was RMB 66.5 billion, also down 6%.
Excluding the contribution of new AI products, third-quarter Non-IFRS operating profit was RMB 83.3 billion, up 8% year over year, with an operating margin of 40%, flat with the same period last year. The gap between the two measures is about RMB 15.2 billion. 'New AI products' mainly include Hunyuan, Yuanbao, CodeBuddy, WorkBuddy, and Xiaowei.
In other words, what is dragging down profit is the AI business, which is still in its investment phase.
Tencent's total capital expenditure in the third quarter reached RMB 81.4 billion, up 527% year over year, with single-quarter spending already exceeding the full-year 2025 total; cumulative capital expenditure for the first three quarters was RMB 166.2 billion, up 179% year over year, with the money flowing mainly into IT infrastructure, data centers, and land use rights. Capital expenditure payments on a cash basis were RMB 65.1 billion, up 226% year over year. R&D investment was RMB 32.36 billion, up 42% year over year; general and administrative expenses rose 29% year over year to RMB 44.1 billion, but excluding R&D expenses, this item grew only 2%, showing that control over non-AI spending remains tight.
Third-quarter free cash flow was negative RMB 69.1 billion. Net cash generated from operating activities was only RMB 2.6 billion, while payments for capital expenditure, media content, and lease liabilities totaled RMB 71.7 billion. Tencent explained that operating cash flow included large computing power purchase prepayments, mainly for Hunyuan upgrades, WorkBuddy inference, and cloud customer demand; excluding these prepayments, free cash flow would have been positive RMB 16.9 billion.
The funding picture changed accordingly: at the end of June there was RMB 58.2 billion in net cash, which by the end of September had become RMB 13.7 billion in net debt. The official explanation is that this was mainly caused by large computing power purchases, only partly offset by RMB 8.4 billion of net inflows from disposals exceeding new investments in the portfolio. Tencent still had RMB 525.7 billion in cash and time deposits on its books, and held listed and unlisted investments totaling about RMB 902.6 billion.
In terms of shareholder returns, Tencent repurchased about 15.87 million shares in the third quarter, spending about HKD 7 billion, and the repurchased shares have been canceled.
There is considerable divergence in the market. Several brokerages cut their target prices ahead of the results, and institutions such as Nomura and UBS had long warned about the pressure of AI spending on margins; but most major banks, including Citigroup, Bank of America, and CLSA, maintained 'Buy' ratings, on the reasoning that near-term profit pressure is consolidating long-term competitiveness. This is essentially a question of time: how long investors are willing to wait depends on how quickly returns materialize.
II. WorkBuddy at a US$500 million annualized run rate and Yuanbao's spending cuts
On the earnings call, Liu Chiping (Martin Lau) disclosed WorkBuddy's revenue figures. Based on the past seven days' revenue, WorkBuddy's annualized revenue has exceeded US$500 million, more than twice Tencent Meeting's revenue; third-quarter revenue grew more than fivefold quarter over quarter from the second quarter.
On the user side, WorkBuddy's daily active users have surpassed 10 million, consuming more than 100 trillion tokens per day, making it the No. 1 AI application in China by large language model token consumption.
In enterprise penetration, the enterprise version was launched in June this year; 60% of China's top 100 companies have become customers, coverage of China's top 500 enterprises has reached 50%, and the product has been deployed in more than 50 industries.
Regarding this monetization path, Lau compared WorkBuddy to the evolution of online games: a compound payment model of subscription plus token usage lets users enter with a low barrier and expand spending as the value they obtain increases. 'When consumption can expand upward along the demand curve, it unlocks a much larger revenue pool.' To market concerns that lightweight flash models would drag down token pricing, his response was that flash is a customer acquisition tool while stronger models handle value monetization, and the two happen to cover different positions on the demand curve.
On the model foundation, Hunyuan has maintained a major version iteration roughly every two months this year. Released on November 4, Hy4 ranked first among Chinese models and first among open-source models globally on Agent Arena at launch, with its hallucination rate dropping from 12.95% to 7.2%. It introduced vision understanding and computer operation capabilities for the first time, focusing on strengthening high-value scenarios such as programming, enterprise workflows, and scientific research.
Hy4 flash, launched this Monday, has 224B total parameters and only 12B active parameters, with a per-task cost about one-twentieth that of Hy4. On its first day online, token calls exceeded 50 trillion, topping the OpenRouter daily leaderboard. On the day Hy4 debuted in WorkBuddy, many users hit their quotas, and a 10-fold capacity expansion was promptly completed — the synergy between model and product is already very tight.
On the cloud business, third-quarter cloud services revenue growth jumped to 35% year over year from 21% in the previous quarter. Within that, model-as-a-service revenue grew more than threefold quarter over quarter, GPU rental revenue rose over 30% year over year, and both WorkBuddy and MaaS revenue achieved more than threefold quarter-over-quarter growth. Management expects this higher growth rate to continue throughout 2026.
The imagination for the consumer side lies within WeChat. Liu disclosed that Xiaowei, the AI assistant embedded in WeChat, has reached 30 million daily active users, and the next step is to integrate it with more than 30 touchpoints including Moments, Official Accounts, and Channels.
He gave a cost figure: even if Xiaowei were opened to all 1.4 billion WeChat users, its quarterly operating cost would still be lower than Yuanbao's quarterly cost at the beginning of this year, because it runs on a lightweight model customized for low cost.
At the same time, given that computing power efficiency multiplies year over year, he expects Xiaowei's capabilities to leap significantly within 12 to 18 months, eventually reaching the level of Meta's products while covering all WeChat users.
In terms of commercialization, Xiaowei will become an amplifier of WeChat's existing revenue, accelerating revenue growth. 'Adding 10% to 20% on top of WeChat's existing revenue is already a very considerable number, and the path will become clearer in the next 6–12 months.' Advertising and transaction monetization are near-term opportunities, with additional revenue from intelligent agent matchmaking services possible later.
As for Yuanbao, Liu said, 'Yuanbao has not been as successful as we expected.' Investment in Yuanbao fell short of expectations, spending has been substantially cut, and the focus has shifted to capability building and the launch of AI search.
Of course, the US$500 million is an annualized figure based on seven days of revenue, and in absolute terms it is still small compared with Tencent's overall scale; the AI business as a whole remains in its investment phase. Whether this new revenue can cover costs will need continued verification over the next few quarters. But at least the 'model–product–revenue' flywheel is already turning.
Beyond the financial figures, Tencent's management did not shy away from the hardest parts of this transformation, either on the earnings call or in internal remarks.
First is the hard constraint on computing power. Faced with analysts' questions, Liu admitted that both leased and self-built computing power face constraints, model training will inevitably be affected, and the release of the next-generation Hunyuan may take longer in the near term.
He also distinguished two stages: inference is currently unaffected, and computing efficiency is improving rapidly year over year; over the long term, the domestic supply chain will gradually enter the inference stage, and the pace of localization will accelerate. This means Tencent must arrange its training cadence in a computing environment with a ceiling, and the timeline for model iteration is now genuinely constrained by external conditions.
On the evening of November 11, Tencent's 28th anniversary, Pony Ma's remarks at an internal LinkTime event quickly leaked, with a severity of tone rarely seen in recent years. He likened the current situation to 'the wave has already come crashing in,' with distant swells swallowing hill after hill: 'You can't still be sitting back admiring it.' He called on core businesses to shift 'from land warfare to sea warfare,' with half the people guarding existing businesses like an army and the other half exploring new AI Native battlegrounds like a navy.
Viewed over a longer timeline — in 2023, Pony Ma's keyword for AI was still 'no hurry,' Hunyuan was released later than peers, and Yuanbao went a long time without being promoted. In January this year, he already reflected at the annual meeting that 'we moved too slowly.' Tencent then launched a series of intensive adjustments: former OpenAI researcher Yao Shunyu became chief AI scientist, overseeing Hunyuan R&D; new departments such as AI Infra and AI Data were set up; in March 2026, the decade-old AI Lab was dissolved and folded into the Hunyuan system, with computing power centrally allocated by headquarters; the Hunyuan team compressed its reporting lines from three levels to two, WXG piloted a project-lead system, and the game publishing line downplayed director and team-lead titles.
On the call, Ma mentioned two examples he was most satisfied with: the model team was rebuilt entirely from scratch and brought Hunyuan to a top position in China within a year; the WorkBuddy team had almost nothing six or seven months ago and built today's business. The ideal state he wants is: 'If we can run existing businesses with fewer people and redeploy those people to build new growth drivers, that is the ideal state.'
This is also the core logic of Tencent's internal remarks a week earlier. Unlike the previous round of passive cost cutting, this one is defined as 'proactive efficiency enhancement': flattening the organization, using small teams as combat units, reducing middle-layer message-passing and approval steps, explicitly stating that 'there may not need to be so many management positions in the future,' and breaking down factionalism and territorial mindsets.
As AI drastically lowers the barrier to development, a small team with digital employees can accomplish work that previously required an entire collaboration chain, and traditional pyramid-shaped management hierarchies are losing their reason to exist.
On the day the results were released, Tencent's shares closed at HKD 372.8, down about 37% year to date. The results did not stop the decline; in the next morning's session, the stock fell in response, at one point dropping below HKD 350.
Judging by this quarter's numbers, Tencent's AI narrative has left behind the light-R&D, fast-iteration model familiar from internet companies of the past and moved toward heavy computing power, long cycles, and ecosystem building. The RMB 81.4 billion in capital expenditure and RMB 69.1 billion of negative cash flow are a form of growing pains at this stage. Whether WorkBuddy can grow into a genuine productivity platform, whether Xiaowei can bring WeChat into the AI era, and whether Hunyuan can hold its place in the top tier amid constrained computing power will all be answered in the coming quarters.
本文来源:https://www.geekpark.net/news/374302
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
